Landscape blog, NOT a peer-reviewed source, and written by a Jolt Atlas author -- so its head-to-head claims favour Jolt Atlas and should be attributed, not stated as fact. Useful for: (1) the overhead-progression framing "1,000,000x -> 100,000x -> 10,000x"; (2) secondhand ZKTorch LLM numbers (GPT-J 20 min, GPT-2 10 min on 64 threads); (3) its own characterisations -- Jolt Atlas "4x-7x faster than anything prior (all before using GPU)", DeepProve "54-158x faster than EZKL" and "not likely privacy preserving", EZKL "65x faster than RISC Zero". Numbers taken from here are tagged provenance: "ICME blog (partisan)".